LREC 2008 Marrakech, Morocco 28 th May 2008 Some Fine Points of Hybrid Natural Language Processing Peter Adolphs, DFKI GmbH, Language Technology Lab, Berlin Stephan Oepen, Universitetet i Oslo, Department of Informatics Ulrich Callmeier, acrolinx GmbH, Berlin Berthold Crysmann, Universität Bonn Dan Flickinger, Stanford University, CSLI Bernd Kiefer, DFKI GmbH, Language Technology Lab, Saarbrücken
Motivation ● hybrid processing, integrating annotations of ‘shallow’ tools into HPSG parsing ● different tools make different assumptions ● example: PTB-style tokenizers for English – e.g.: Don't you! → <do, n't, you, !> – contracted verb forms are split – punctuation is split off the preceding word form ● we need to adapt annotations of different tools to the requirements of our grammar ● goal: a declarative, expressive, scalable device
Token Feature Structures ● feature structures for describing tokens ● different annotations provided as feature structures ● lattice of structured categories (token feature structures) as input to the parser
Generalized Chart ● tools may assume different tokenization (paradigm case: input from speech recognizers) ● chart: dag whose vertices are abstract objects rather than indexed token boundary positions
Chart Mapping ● chart mapping: non-monotonic rewrite mechanism on feature structure chart edges ● general format: [ CONTEXT : ] INPUT → OUTPUT ● CONTEXT, INPUT, OUTPUT are sequences of feature structures (each possibly empty) ● resource-sensitive: chart edges that let a rule fire may be removed (namely, all INPUT edges)
Chart Mapping – Example ● example: recombining split contracted forms ● rules extended with regular expression matches ● regex capture groups can be referred to in the output ● rules themselves described as feature structures, thus we can use re-entrancies
Chart Mapping – Examples ● light-weight named entity recognition ● fixing broken tokenization
Previous Architecture (Simplified) ● preprocessing has to provide natural language input the input chart as expected by the grammar Preprocessing ● this has to be ensured by specialized conversion routines without recourse to the grammar Lexical Instantiation ● changes to the grammar have to be reflected in these Syntactic Parsing data adaptation routines SYN ... SEM ...
Proposed Architecture (Simplified) ● proposal: token mapping per- natural language input forms certain preprocessing Preprocessing steps within the grammar ● advantages: – full control for the grammar Token Mapping writer, using the same formalism as for the grammar Lexical Instantiation – makes assumptions by the grammar explicit Syntactic Parsing – removes complexity from preprocessing SYN ... SEM ...
Hybrid Processing ● shaping the search ● constraints on the space of the parser: search space – widening search – hard: categorial space (e.g. unknown conditions for word handling) introduction / removal of chart edges – narrowing search – soft: probabilistic space (e.g. removing / postponing the disambiguation, processing of edges) prioritize parser's tasks on the agenda
Lexical Instantiation ● native and generic lexical entries (les) ● selection of appropriate generic lexical entries originally controlled by the parser (hard-coded) ● strategy: – map from part-of-speech tags to generic les – instantiate generic le for highest ranked pos tag where no native le is available ● disadvantage: – not flexible enough (e.g. no chain of responsibility) – partial lexical coverage: We’ll bus to Paris.
Lexical Instantiation ● proposal: try to instantiate all generic les for all tokens ● token feature structure is unified into a predefined path in the lexical entry ● selection of compatible tokens by constraints on the token feature structure ● example:
Lexical Filtering ● after lexical instantiation, native and generic les may be available in the same chart cell ● we can restrict lexical instantiation by positing constraints on the token feature structures ● but we might also want to prevent some lexical chart edges in certain contexts (set operations) ● proposal: lexical filtering phase ● same formalism as for token mapping: chart mapping rules with empty OUTPUT list
Proposed Architecture ● use feature structures to natural language input describe tokens Preprocessing ● chart mapping: resource- Token Mapping sensitive rewriting of feature structure items Lexical Instantiation ● chart mapping on token fs Lexical Parsing ● generic instantiation driven by Lexical Filtering compatibility with token fs Syntactic Parsing ● lexical filtering with chart mapping SYN ... SEM ...
Applications ● fine grained control over instantiation of generic lexical entries ● mapping external morphological information into the grammar's universe ● chart dependency filter (optimizing parsing performance) ● activate syntactic rules only for certain spans of the input (e.g., in hybrid grammar checking)
Conclusions ● versatile device for many applications ● external information is made accessible to the grammar ● pre-processing can be better controlled with grammar-specific means ● reduces the need for special code inside and outside the parser ● outlook: consilidation of our current parsers and grammars
Thank you!
Acknowledgements ● DELPH-IN community and beyond, especially Nuria Bertomeu, Ann Copestake, Remy Sanouillet, Ulrich Schäfer and Benjamin Waldron for numerous in-depth discussions ● funding: – ProFIT program of the German federal state of Berlin and the EFRE program of the EU (to the DFKI project Checkpoint) – the University of Oslo (through its scientific partnership with CSLI)
Recommend
More recommend
